mujib-llm-lora

mujib-llm-lora

Pashto-focused LoRA fine-tuned Qwen2.5 7B model trained with Unsloth for efficient and fast inference.

Model Information

• Model Name: mujib-llm-lora
• Developed by: mhalimi3008
• License: Apache-2.0
• Finetuned from model: unsloth/qwen2.5-7b-unsloth-bnb-4bit

This Qwen2 model was trained 2x faster with Unsloth.

---

Overview

mujib-llm-lora is a Pashto language instruction-tuned model designed for:

• Pashto conversations
• Question answering
• Text generation
• Translation
• Educational assistance
• General NLP research

The model is optimized using:

• LoRA fine-tuning
• 4-bit quantization
• Unsloth acceleration
• PEFT optimization

---

Features

• Fast inference
• Low VRAM usage
• Optimized for consumer GPUs
• Pashto language support
• Instruction-following capability
• Efficient 4-bit loading

---

Installation

# =====================================================
# Install dependencies
# =====================================================
!pip install -q unsloth transformers accelerate peft bitsandbytes

---

Example Usage

# =====================================================
# Imports
# =====================================================
import torch
from unsloth import FastLanguageModel
from transformers import TextStreamer

# =====================================================
# Model Names
# =====================================================
base_model = "unsloth/qwen2.5-7b-unsloth-bnb-4bit"
lora_model = "mhalimi3008/mujib-llm-lora"

# =====================================================
# Load Model + Tokenizer
# =====================================================
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = lora_model,
    max_seq_length = 2048,
    dtype = None,
    load_in_4bit = True,
)

# Enable faster inference
FastLanguageModel.for_inference(model)

# =====================================================
# Test Prompt
# =====================================================
prompt = """### Instruction:
په پښتو ژبه خپل ځان معرفي کړه.

### Response:
"""

inputs = tokenizer(
    [prompt],
    return_tensors="pt"
).to("cuda")

# =====================================================
# Generate Response
# =====================================================
text_streamer = TextStreamer(tokenizer)

outputs = model.generate(
    **inputs,
    streamer=text_streamer,
    max_new_tokens=200,
    temperature=0.7,
    top_p=0.9,
    do_sample=True,
)

# =====================================================
# Decode Final Output
# =====================================================
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

print("\n\n========== FINAL RESPONSE ==========\n")
print(response)

---

Example Output

سلام! زه یو مصنوعي ذهانت ماډل یم چې په پښتو ژبه خبرې کولی شم او ستاسو پوښتنو ته ځوابونه درکوم.

---

Training Details

Base Model

unsloth/qwen2.5-7b-unsloth-bnb-4bit

Training Method

• LoRA fine-tuning
• PEFT
• 4-bit quantization
• Unsloth optimized training

---

Intended Use

This model is intended for:

• Pashto AI assistants
• Chatbots
• Research
• Educational systems
• NLP experimentation
• Translation systems

---

Limitations

• The model may generate inaccurate information.
• Responses may occasionally mix languages.
• Performance depends on dataset quality and coverage.
• Human verification is recommended for important tasks.

---

Hardware Requirements

Recommended:

• NVIDIA GPU
• CUDA support
• 12GB+ VRAM recommended
• Python 3.10+

---

Libraries Used

• Transformers
• Unsloth
• PEFT
• Accelerate
• BitsAndBytes
• PyTorch

---

Citation

@misc{mujib_llm_lora_2026,
  author       = {mhalimi3008},
  title        = {mujib-llm-lora},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/mhalimi3008/mujib-llm-lora}}
}

---

Acknowledgements

Special thanks to:

• Unsloth
• Qwen Team
• Hugging Face
• Transformers Library
• PEFT Library

---